An architecture of high performance cluster network: Maestro2

Cluster computers have become the platform of choice to run high performance parallel computing applications. However, most cluster computers use conventional wide-area network technologies to interconnect the individual processing elements. Therefore, there is a disparity between the wide-area network technologies used and the network technology required by cluster computing, leading to a severe degradation in performance. Maestro network technology attempts to solve this disparity. This paper describes improvements to the basic Maestro architecture. In particular it describes a new technique used by Maestro2 (continuous network burst) and shows its impact in the overall performance of Maestro cluster networks.